The following graph shows the results of this scaled-problem test:
The second test is for a fixed problem size of 3.2M grid points using seven levels of refinement. The mesh is distributed as evenly as possible across the processors, so that the number of 8x8x8 grid blocks decreases from 49 per PE on the 128-PE partition down to 12 per PE on the 512-PE partition. This last run attained 9.9 GFlops, or 220 times the C90 performance. The per-PE performance on the T3E deteriorates noticeably as the number of PEs rises, with the problem size fixed, because the ratio of time spent communicating to that doing useful work increases significantly.
The following graph shows the results of this fixed-problem test:
The package of AMR routines with which the NRL MHD code was merged was written by Peter MacNeice, Kevin Olson, Clark Mobarry, Rosalinda de Fainchtein, and Charles Packer. Many helpful discussions with them during a cooperative development effort on this code are gratefully acknowledged. Thomas J. Clune of Cray also provided a number of helpful suggestions and ran several early test cases with the code-in-progress, for which he deserves thanks.